24 June 2026 in SSRN Electronic Journal
Phase reconstruction from a given amplitude spectrogram is a fundamental but ill-posed problem in speech and audio signal processing. Conventional analytical methods explicitly enforce amplitude consistency but lack a learned speech prior, often producing perceptual artifacts. Diffusion-based methods can model natural speech distributions, but the target amplitude is usually used only as a soft condition and is not explicitly satisfied during sampling. To address this limitation, we propose Amplitude-Projected Diffusion Posterior Sampling (AP-DPS) for speech phase reconstruction. AP-DPS is based on an amplitude-conditioned deterministic diffusion model. At each reverse step, the network predicts a clean waveform candidate, which is projected onto the target amplitude spectrogram by replacing its STFT amplitude with the given amplitude while preserving the estimated phase. The projected clean estimate is then used in the DDIM update. AP-DPS can also incorporate an additional measurement through a DPS-style measurement gradient during inference, without retraining the diffusion model. Experiments on the VCTK corpus show that AP-DPS improves phase reconstruction under both measurement-free and measurement-guided settings. With an additional measurement obtained from a signal sampled at 1200 Hz, all compared methods improve across PESQ, eSTOI, and SNR. AP-DPS achieves the best PESQ in both settings, while diffusion-based methods show higher SNR than analytical methods. However, analytical methods remain stronger in eSTOI. Zero-shot experiments further demonstrate that AP-DPS can exploit low-sampling-rate and noisy-waveform measurements without task-specific retraining.